12.00_2014_OFA_Workshop_iWARP_Updatex

download report

Transcript 12.00_2014_OFA_Workshop_iWARP_Updatex

iWARP Update
#OFADevWorkshop
Increasing interest in iWARP
• RDMA Use Cases
–
–
–
–
–
–
High Performance Computing
File and Block Storage
NVM access
Virtual Machine migration
Low-latency messaging middleware
Virtualization and Cloud deployments place important requirements
on these use cases
• Key iWARP value propositions for these use cases
–
–
–
–
Engineered for “typical” Ethernet (best effort, no DCB, no QCN, etc)
Natively Routable
Multi-pathing supported at Layer 3 (as well as Layer 2)
Reliable and proven TCP Transport
• Mature and efficient retransmission algorithms
• Dynamic and verified congestion algorithms
March 30 – April 2, 2014
#OFADevWorkshop
2
iWARP Standards
• iWARP updates and enhancements are handled
by the IETF STORM (Storage Maintenance)
working group
• Finalized RFCs
RFC 5040
RFC 5041
RFC 5044
RFC 6580
RFC 6581
March 30 – April 2, 2014
A Remote Direct Memory Access Protocol Specification
Direct Data Placement over Reliable Transports
Marker PDU Aligned Framing for TCP Specification
IANA Registries for the RDDP Protocols
Enhanced RDMA Connection Establishment
#OFADevWorkshop
3
iWARP In-Progress RFCs
• draft-ietf-storm-rdmap-ext-09.txt
– Extends RFC 5040
• Adds Atomic Operations and Immediate Data
– Authors from Intel, Broadcom, Chelsio
– IESG approved. Next step RFC Editor Queue
• draft-ietf-wood-rdmap-ext-v2-00.txt
– Extends RFC 5040
• Add Send with Immediate Data
• Add IB-style RDMA Read
– Authors from Intel
– Submitted for initial review by STORM working group
March 30 – April 2, 2014
#OFADevWorkshop
4
Traditional iWARP RDMA Read
SQ WQE
RDMA Read op
STag describing Sink buffer
STag describing Source buffer
DMA
Read
Sink
Buffer
RDMA
NIC
RDMA
NIC
Source
Buffer
DMA Write
SQ WQE
Local Invlidate op
STag describing Sink buffer
This slide has animations – view in Presentation mode
March 30 – April 2, 2014
#OFADevWorkshop
5
IB-style RDMA Read
SQ WQE
RDMA Read op
SGL describing Sink buffers
R_Key or STag describing Source buffer
DMA
Read
Sink
Sink … Sink
Buffer Buffer
Buffer
RDMA
NIC
RDMA
NIC
Source
Buffer
DMA Writes
No Sink Buffer Local
Invalidate required, because
Sink buffer R_Key (STag)
not exposed to network
In-progress RFCs enable this flow on iWARP
This slide has animations – view in Presentation mode
March 30 – April 2, 2014
#OFADevWorkshop
6
Goal of In-Progress RFCs
• Common Application capabilities across all
flavors of RDMA
• These RFCs remove all known application
differences between iWARP and InfiniBand
March 30 – April 2, 2014
#OFADevWorkshop
7
IETF Alignment/Synergy with iWARP
• iWARP currently leverages:
– TCP
• Reliable transport and congestion management
– Explicit Congestion Notification
• Inherited from TCP/IP layers
• iWARP will naturally adopt/use:
– Tunneling/Network Overlays
• iWARP works with (but does not require) existing tunnel protocols
(ie Generic Routing Encapsulation) and NVO3 technology investigations
• Connectionless messaging to complement iWARP RDMA
– Typically realized with unreliable datagrams (unicast and multicast)
– InfiniBand has Unreliable Datagram (UD)
– UDP may be used in place of UD for Ethernet implementations
• No new wire protocol standards required
March 30 – April 2, 2014
#OFADevWorkshop
8
iWARP Ecosystem
• Strong industry support to evolve iWARP
• Good alignment with IETF, and support in
STORM to evolve the standards
• OFED 3.5-2 stable drivers from multiple
vendors: cxgb3, cxgb4, nes
• Intel is implementing iWARP RDMA as a key
capability in Fort Park. Fort Park is an
Ethernet IP block that will be integrated into
future Intel server chipsets.
March 30 – April 2, 2014
#OFADevWorkshop
9
Call to Action
• Participate in STORM standards reviews
• iWARP RNIC vendors and system software
vendors consider supporting the in-progress
RFCs as soon as possible
• Develop future RDMA extensions with a goal to
enable them across all flavors of RDMA
March 30 – April 2, 2014
#OFADevWorkshop
10
Thank You
#OFADevWorkshop